EXPERT SYSTEM BASED ON OBJECT-ORIENTED APPROACH FOR LAND COVER
MAPPING
Zhang Lei, *, Zhou Yueming, Wu Bingfang
Institute of Remote Sensing Applications, Chinese Academy of Sciences, Beijing 100101, China.
- zhang@irsa.ac.cn
Commission VII, WG VII/4
KEY WORDS: Expert classification, Object-oriented classifier, land cover, decision tree, segmentation
ABSTRACT:
The hybrid classifier is presented in combination of expert system and object oriented approach, for which increased information is
added for classification and improve the accuracy. Instead of the original image bands, derived data are prepared for classification,
which contains physical meaning and clear separation of recognition and assessment of object classification, the variables of NDVI,
seasonal change vegetation index, vegetation brightness index, hard surface brightness index, moisture stress index are derived from
raw image data. The 17 classes of land cover are possible for classification and make high overall accuracy of 86%, the classes of
evergreen needle leaf, evergreen shrub, sands/beach, lake/reservoir get high accuracy in classification, the classes of grass, grass
dominated urban vegetation, bare land/rocks have poor classification, the influence to classification is the factors of DEM, slope,
shadow. Expert system based on object orient approach has many advantages of multi-spectral cluster separation and object analysis
for land cover classification, especially for more than 10 land cover classes. Multiple variables are better solution in criteria setup
than single variable and reduce the hierarchical levels.
1. INTRODUCTION
With increasing concerns on global environment, climate
change and bio-diversity, land cover/use researches get great
progresses in the recent decades, instead of the traditional
approaches of classification for land cover, the new
methodologies are developed, such as fuzzy logic (Foody 1996),
decision tree (Breiman et al., 1984; Hansen et al. 1996), neural
network (Carpenter et al. 1997), support vector machine (Chang,
2001), expert system (William, 2001), object oriented
classification (Geneletti, 2003) and hybrid model (van der Meer
1995). Each classification strategy has its advantage and
limitation, the univariate decision tree is not based on any
assumptions of normality within training statistics and is well
suited to situations where a single cover type is resented by
more than one cluster in the spectral space, the decision tree
yields a set of rules which are easy to interpret for an expert
analyst in correcting splits associated with faulty or
contradictory training data (Hansen, 2000). neural network is
also independent from the data prior probability distribution and
capable of self-study, self-adaptive and improve the
classification accuracy 10-20% high over traditional classifiers,
suitable for complicate classification (Richard, 2001); object
oriented approach is taken on the vector domain information
extraction of spatial geometry, spatial relations, suitable for
high resolution image application, Expert systems allow the
expert establish the logical decision tree based on feature space
and class physical features for the universal application in other
places, and integration of remotely sensed data with other
sources of geo-referenced information such as land use data,
spatial texture, and digital elevation models (DEMs) to obtain
greater classification accuracy (William, 2001).
Nevertheless, the classification of land cover on meso-scale is
still at low accuracy for large area due to the spectral mixture
among classes, spatial heterogeneity and image contamination,
the strategy of multiple classifier combination is considered as
an approach for accuracy improvement, this paper is discussed
on the combination classifiers of object oriented classifier and
expert system for land cover classification. Object oriented
classifier is applied for derivation of the contextual and spatial
geometric information of classes from the images, especially for
application on high resolution and meso-scale resolution
classification, expert system is allowed to other ancillary data
sources adding to classification or post-classification, such as
DEM, slope which are more relative to spatial distribution of
land cover, with its strategy of hierarchical structures of
decision tree, it is not based on any as sumptions of normality
within training statistics and is well suited to situations where a
single cover type is resented by more than one cluster in the
spectral space, decision trees yield a set of rules which are easy
to interpret and suitable for deriving a physical understanding
of the classification process.
The study area is selected in the mountain area in Three Gorge
area, Chongqin, China. It occupies approximately 100 km2 in
mid-latitude warm temperate zone with dominated natural
vegetation of evergreen shrub and deciduous broad leaf forest.
With high density of population, the agricultural land is major
landscape of land cover. SPOT5 XS and Pan data are acquired
between 2004-2006 for cloud free images, two scenes of
different seasons for each area is selected for vegetation
detection, one time scene for leaf on date is selected in July or
August, which vegetation is grow up most, another scene for
leaf off data is selected in January or February, which most
deciduous vegetation is dormant, for the confusion of cropland
* The paper is supported by the project of ‘real-time monitoring with remote sensing for eco-environment in Three Gorge Dam’ (No.
sx[2002]-004). Corresponding author: Wu Bangfang, Institute of Remote Sensing Applications, CAS
679
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
with natural vegetation, the phenology of crops are considered
in acquisition selection, one of scens must be in harvest and
another season in growing up. The top map of 1:50, 000 is
come from Bureau of State Surveying and Mapping for
derivation of DEM and slope data.
SCVI =
NDVI LON − NDVI LOFF
NDVI LON + NDVI LOFF
……….(Eq. 3)
NDVI LON : maximum NDVI year round, normal in Month 7-8
NDVILOFF : minimum NDVI year round, normal in Month 1-2
2. METHEDOLOGY
2.1 Data pre-processing
data are regarded as environment factors affected indirectly on
the land cover, include the DEM, Slope, Aspect, DEM and
slope are applied in this classification due to their high relative
to the vegetation spatial distribution. The geometric parameters
is obtained from object oriented approach to derive the object
parameters, such as ratio of length/width, ratio of perimeter/area,
roundness, spatial relations, object hierarchical relationship of
multi-scale.
The images are corrected for the effects of atmosphere, and
converted to calibrated reflectance using commercially
available software that incorporates the MODTRAN3 radiative
transfer code (ATCOR2 for the ERDAS Imagine software;
GEOSYSTEMS
GmbH,
1997).
A
middle-atitude
summer/winter, rural aerosol concentration model therefore was
used as input to the radiative transfer code, estimated visibility
is input from ground climate station at the same day. SPOT XS
and Pan data are fused to 2.5 m grid with the approach of High
Pass Filter (HPF), which keep least spectral change during the
fusion processing, All coverages were geo-referenced to the
Beijing 54, Gauss Kruger Zone 108 or 111, coordinate system.
2.2 Classification system
FAO land cover classification system (LCCS) is accepted in
this study, LCCS is comprehensive, standardized a priori
classification system, independent of the scale or means used to
map. It enables a comparison of land cover classes regardless of
data source, economic sector or country (Antonio, 2004). There
are 17 classes of land cover are found in study area (table 1).
The radiometric influence of topography on recorded sensor
signals has distorted the vegetation features and physical
structure. This can lead to image classification error as well as
error in forest classification and parameter estimates, The
topographic calibration on reflectance values is necessary in the
mountain area (Equation 1,2).
L(λ , e ) = Ln ⋅ (λ ) ⋅ cos i
Table 1 FAO land cover classes in Three Gorge area
1. evergreen broad leaf
2. evergreen needle leaf
3. evergreen shrub
4. deciduous shrub
5. grass
6. tree dominated urban vegetation
7. tree and shrub dominated urban vegetation
8. grass dominated urban vegetation
9. herbaceous plantation (upland)
10. aquatic graminoids (paddy)
11. urban area
12. industry
13. sands/beach
14. bare land/rocks
15. river
16. pond
17. lake/reservoir
(Eq. 1)
L ：solar radiance
λ ：bands
e ：ground slope
Ln ：corrected solar radiance
i ：incident angle
cos i = cos (90 − θ s ) ⋅ cos θ n + sin (90 − θ s ) ⋅ sin θ s ⋅ cos (φ s − φ n )
( Eq. 2)
2.3 Hybrid classification
θ s ：solar elevation angle
θ n ：ground slope
φ s ：solar azimuth angle
φ n ：ground azimuth angle
Instead of the original image bands, derived data are prepared
for classification, which contains physical meaning and clear
separation of recognition and assessment of object classification,
in this study, NDVI, seasonal change vegetation index (SCVI)
(Equation 3), vegetation brightness index (VBI: Nir+Swir), hard
surface brightness index (HBI: Green+Red), moisture stress
index (MSI: Nir/Swir) are produced. The ancillary
680
The hybrid classification is based on the combination of object
oriented approach and expert system classification. The
approach of expert system has the strategies of non-parameter
separation and multi-variable decision rules, Use of the expert
system has an advantage to classify multiple clusters of feature
space for one class, such as upland and sands/beach, which is
not allowed to operate in maximum likelihood classification
(MLC) approach ， expert system has higher accuracy of
classification than MLC in complicate spectral classes; The
object oriented appoach is focused on the extraction of spatial
geometry information, some artificial classes have similar
spectral features, but have different geometric and relation
properties, which are helpful to recognize and divide, such as
reservoir/pond, industry and urban vegetation.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
raster types to extract both object geometric and spectral
information, each object spectral information includes average
value, standard deviation, maximum value, minimum value of
inner pixels among each band, each object geometry includes
the features of area, size, spatial relation to other objects. These
integrated data are input to the database of expert knowledge.
According to the object classified rules, the top down decision
trees are set up for classification. The selection of physical
meaning classification rules and ancillary data are considered to
ensure the reliable result and stable classified procedure, 13
decision trees with one or multiple criteria are made for
classification of 17 classes (Figure 2). The design of expert
system is based on the image spectrum, object properties and
ancillary data features, the maximum differences of features are
considered on upper hierarchical steps to reduce the omission
errors. According to the analysis of samples on the image,
upper step is to classify three groups of water-body, vegetation
and non-vegetation, then classify classes in each groups.
The hybrid classifier model was constructed using the Definiens
5.0 image processing software. The classification is performed
in object oriented method in advance and then the expert system
followed (Fig. 1), that compose of 4 major steps of
segmentation, object creation, expert knowledge
Figure1 follow chart for hybrid classification of land cover
rule setup and classification. The segmentation is based on the
unsupervised classification to form multiple scale cluster layers,
for this study, 10 and 20 scales are selected to shape different
size cluster, most parts of classed are segmented in 10 scale, but
some classes with unique spectrum or the phenomenon of salt
and pepper are segmented in 20 scale, such as water body and
artificial classes, the purpose of scaling segmentation is to form
best shape for the representation of object properties. vector
objects are then transformed from the segmented clusters of
Figure 2 the classification rules and steps of expert knowledge system
residence, sands/beach, pond lake/reservoir, for which the
accuracy are also more than 88%. The classes with both low
omission and commission are evergreen needle leaf, evergreen
shrub, sands/beach, lake/reservoir which have least errors in the
classification. Big errors are occurred in the classes of grass,
grass dominated urban vegetation, bare land/rocks, for which
accuracy has about 45%.
3. RESULTS AND ANALYSIS
The accuracy assessment of classification was measured using
randomly selected points for which land cover was determined
using an orthophoto mosaic geo-referenced to the image. The
product accuracy, user accuracy, overall accuracy are analyzed
from the sample error matrix assessment (Table 2). Overall
classification accuracies for the 17 classes reach 86%, for
product accuracy, the classes with little omission numbers are
evergreen needle leaf, evergreen shrub, tree dominated urban
vegetation, upland, industry, sands/beach, lake/reservoir, for
which the accuracy are more than 88%. For user accuracy, the
classes with least commission numbers are the classes of
evergreen broad leaf, evergreen needle leaf, deciduous shrub,
The water-body is easy to separate with the two variables of
winter VBI and NDVI when chlorophyll content is little in the
water, but the hill shadows which have same spectral feature are
larger than in the summer, and result in some errors. Then river,
reservoir, pond are divided based on the object orient approach
which are impossible classified with MLC. The groups of
681
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
Breiman, L., Friedman, J. H., Olshend, R. A., and Stone, C.
J.1984, Classification and Regression Trees, Wadsworth,
Monterey, CA.
natural vegetation, cropland and non-vegetation are at the same
level for classification with two variables of winter NDVI and
summer NDVI, the rules are two high NDVI values for the class
of vegetation, one high NDVI for cropland and two low NDVI
for non-vegetation, non-vegetation obtains high accuracy
classification, but cropland is mixed with natural vegetation,
because some crops during harvested season cover with natural
grass instead of bare land, other reason is local elevation
difference resulting in the crop phenological change, and
spectral heterogeneity produces some errors in classification.
SCVI is better variable for separation of deciduous and
evergreen vegetation, VBI is more sensitive to vegetation type
than NDVI and MSI in the classification of tree, shrub, and
grass. Urban vegetation classification is based on the rule of
spatial relation of closeness to the residence, the variable of
relative border ratio is used, some outside natural vegetation is
classified to urban vegetation, and some inner urban vegetation
object is omitted in classification.
Carpenter, G. A. Gjaja,M. N., Gopal, S., and Woodcock, C.
E.,1997, ART neural networks for remote sensing: vegetation
classi. cation from Landsat TM and terrain data. IEEE T
ransactions on Geoscience and Remote Sensing, 35, 308–325.
Chang C.C., Lin C.J.2001, Training v-support vector classifiers:
theory and algorithms, Neural Computation Vol. 13(9), pp.
2119-2147.
Foody, G. M.,1996, Approaches for the production and
evaluation of fuzzy land cover classi. cations from remotelysensed data. International Journal of Remote Sensing, 17, 1317–
1340.
Geneletti, D, GORTE B. G. H.,2003, A method for objectoriented land cover classification combining Landsat TM data
and aerial photographs, International Journal of Remote
Sensing, Vol.24, No.6,1273-1286.
4. CONCLUSIONS AND DISCUSSIONS
Expert system based on object orient approach has many
advantages of multi-spectral cluster separation and object
analysis for land cover classification, especially for more than
10 land cover classes. Multiple variables are better solution in
criteria setup than single variable and reduce the hierarchical
levels. The 17 classes of land cover are possible for
classification and make high overall accuracy of 86%, the
classes of evergreen needle leaf, evergreen shrub, sands/beach,
lake/reservoir get high accuracy in classification, the classes of
grass, grass dominated urban vegetation, bare land/rocks have
poor classification, the influence to classification is the factors
of DEM, slope, shadow. Some classes which are difficult to
classify in MLC are available in hybrid classifier, such as urban
vegetation from vegetation, reservoir from river.
Hansen, M., Dubayah, R., and DeFries, R.,1996, Classi. cation
trees: an alternative to traditional land cover classi. ers.
International Journal of Remote Sensing, 17, 1075–1081.
Hansen, M., Dubayah, R. S. DEFRIES†, J. R. G.
TOWNSHEND and R. SOHLBERG.,2000, Global land cover
classification at 1 km spatial resolution using a classification
tree approach. International Journal of Remote Sensing, 21,
No6-7, 1331-1364.
Richard O. Duda, PeterE. Etc.2001. Patern Classification.
ISBN:0-471-05669-3.
Van der Meer, F.,1995, Spectral unmixing of Landsat Thematic
Mapper data. International Journal of Remote Sensing, 16,
3189–3194.
However, the rule setup of expert system is complicate, timeconsumed, and requires some experience of land cover features
and spatial distribution.
5. REFERENCES
William L. Stefanov, Michael. Ramsey, Philip R.
Christensen,2001, Monitoring urban land cover change: An
expert system approach to land cover classification of semiarid
to arid urban centers, Remote Sensing of Environment 77, 173–
185
Antonio Di Gregorio, Land Cover Classification System, Food
and Agriculture Organization of the United Nations,
Rome,2004
682
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
type
1
2
3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
total
P
accuracy
Overall
accuracy
797186
18004
127727
3192410
14897
22124
3611747
15094
93345
139094
64208
462743
4701
6513
34487
41298
75690
89669
6179
4206
4
5
6
11595
4204
341893
17011
7
8
12889
63129
23170
2213
28788
10
11
12
13
14
15
16
9307
2107
86898
14809
9587
8912
63698
31597
3613
7690
27596
70122
9
15597
28210
93257
5108
6995
39411
8488
149267
10597
16877
44304
58112
40985
2913
4021139
15021
4497
48098
50405
5800
2136491
467048
756733
46065
1198759
1501
7985
9610
14702
18785
293666
1709016
3458546
4105419
159874
126239
104559
199950
102417
4080172
73082
2696043
775518
1198759
18893
74717
395372
47%
92%
88%
44%
50%
89%
75%
43%
99%
69%
79%
98%
100%
51%
85%
74%
86%
Table 2 Error matrix for hybrid classification of land cover in Three Gorge area
683
U
accuracy
95%
99%
95%
100%
57%
41%
43%
26%
79%
75%
90%
61%
96%
52%
81%
88%
13401
100%
13401
17
100%
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
684